# CMP9764 — Advanced Robotics: Assessment Item 2 # DMP vs SEDS: Trajectory Learning under Single-Demonstration Conditions Implementations of Dynamic Movement Primitives (DMP) and the Stable Estimator of Dynamical Systems (SEDS), compared for learning from a single demonstration across four 2D trajectory datasets of increasing geometric complexity. ## Methods - **DMP:** canonical + transformation system with a forcing term fit by Locally Weighted Regression. - **SEDS:** GMM-based velocity field trained under a Lyapunov stability constraint (global convergence guarantee). - **Principled hyperparameter selection:** BIC, elbow detection, and closed-loop stability verification, with arc-length-MSE evaluation that isolates geometric path fidelity from speed. ## Key findings - DMP gives better path fidelity on 3/4 datasets (up to 15.4x lower arc-MSE on sharp trajectories); SEDS guarantees sub-0.2 mm goal convergence everywhere. - The MSE training objective outperforms log-likelihood by factors of **x17 to x56,399** under data-sparse conditions — reversing the standard multi-demonstration finding. - ## Prerequisites ## Getting started ```bash git clone https://github.com/riceth/dmp_vs_seds_trajectory_learning.git pip install -r requirements.txt ``` ## MATLAB (Stage 1 — SEDS) - MATLAB R2019b or later - `SEDS_lib/` and `GMR_lib/` toolboxes present in the project root ## Python (Stages 2 and 3 — DMP and Comparison) - Python 3.8 or later Install dependencies: ```bash pip install numpy matplotlib scipy pandas scikit-learn pydmps ``` --- ## Setup ## Getting started ```bash git clone https://github.com/riceth/.git pip install -r requirements.txt ``` Ensure the `data/` folder contains all four CSV files. Ensure `SEDS_lib/` and `GMR_lib/` are in the project root (MATLAB will add them to path automatically). --- ## Running the Pipeline ### Stage 1 — SEDS (MATLAB) Open MATLAB, navigate to the project root, and run: ```matlab demo_SEDS_Learning ``` This will: - Run 50 seeds per K value for all four datasets - Select optimal K via BIC (or MSE elbow for Line) - Run stability verification on the top 3 seeds - Compare MSE vs log-likelihood objectives - Save metrics to `output/logs/seds_*.csv` - Save figures to `output/figures/seds_*.png` **Expected runtime:** 10–20 minutes depending on hardware. --- ### Stage 2 — DMP (Python) From the `src/` directory, run: ```bash cd src python dmp_reproduce.py ``` This will: - Train DMP at baseline n\_bfs = 500 on all datasets - Sweep n\_bfs over {25, 50, 100, 150, 175, 200, 250} with three selection criteria - Run qualitative overfit analysis at n\_bfs ∈ {50, 200, 1000, 2000} on WShape and SShape - Save trajectory logs to `output/logs/dmp_*.csv` - Save figures to `output/` **Expected runtime:** Under 2 minutes. --- ### Stage 3 — Comparison (Python) **Must be run after both Stage 1 and Stage 2 are complete.** From the `src/` directory, run: ```bash python comparison_pipeline.py ``` This will: - Compute arc-length position MSE for DMP and SEDS on all four datasets - Generate 8 comparison figures - Save a summary CSV to `output/comparison/summary_comparison.csv` - Print the DMP vs SEDS arc-MSE table to the terminal **Expected runtime:** Under 1 minute. --- ## Notes - Run stages **in order** (SEDS → DMP → Comparison). The comparison pipeline reads logs from both previous stages. - If you re-run the SEDS pipeline, existing CSV logs in `output/logs/` will be overwritten. - The SEDS pipeline suppresses MATLAB singularity warnings during execution. These are expected and handled internally. - All Python scripts must be run from the `src/` directory to ensure relative paths resolve correctly. # AdvbancedRobotics2